top of page
00 SoftUtter Logo 1017x287.png

Spoken Scaffolding for L2 French Learners

LLM-driven conversational repair and low-anxiety oral production.

Role 

Duration

Tools

Participants

AI Prototyper & UX Researcher

Sep 2026 – Present

Python, Streamlit, OpenAI API, Docker, gTTS

Beginner to Intermediate L2 French Learners 

Motivation

The Gap between Learner Intent and Pronunciation

Originating from my Langogo research on conversational repair, SoftUtter is a rapid functional prototype providing multimodal, low-anxiety oral scaffolding through text and synthesized speech without complex interfaces.

01 OV B 1771x553.png

Design Goal

Phonological Scaffolding

I designed an oral feedback interface prioritizing communicative fluency and phonological scaffolding, assisting L2 French learners in acquiring the correct pronunciation of French words.

02 Goal 1703x637.png

Technical Implementation

Development Tools

I developed the web interface using Streamlit in my Python coding process and deployed it via a lightweight Docker container on the hosting platform.

03 Demo 1 1853x1293.png

​The Spoken Recast

The backend integrates the DeepSeek API using structured JSON output mode to separate the spoken recast target, English semantic confirmation, and phonetic production notes into dedicated interface elements.

03 Demo 2 1133x943.png

From Learner Intent to Target Utterance

Audio synthesis is generated dynamically in-memory via gTTS and rendered into the browser through HTML5 audio buffers, avoiding persistent server file overhead.

03 Demo 3 1831x899.png

Testing and Qualitative Evaluation

I tested the system and its affordances, finding that the spoken recast immediately models the spoken target, allowing learners to hear natural liaison flow before speaking.

03 Demo 4 1583x1113.png

Iterating

Add the Audio Input Affordance

Having noticed that many users struggled with typing French words, I added an Audio Input affordance to SoftUtter, offering a convenient, hands-free way for learners to attain the accurate target French utterances.

itr 0 1933x1319.png

Multimodal Input: Audio and Text

This new affordance provides a fast, natural way for L2 French learners to record their voice and instantly generate accurate spoken scaffolding.

itr 1 1821x1027.png

Retaining the Text Input Affordance

Retaining the Text Input affordance remains essential for learners who prefer composing sentences carefully or working in noise-sensitive environments.

itr 2 1815x1113.png

My Takeaways

LLMs' Scaling Effects on Language Learning

Through this project, I realized how coupling prompt-engineered LLMs with lightweight speech synthesis can enable scalable, low-anxiety oral recasting in SLA.

Trade-offs and Limitations

While rapid prototyping provides immediate feedback loops, trade-offs include latency in network round-trips and the lack of fine-grained acoustic analysis.

Future Iterating Directions

Future iterations will incorporate client-side audio capture to evaluate learner shadowing attempts against the generated acoustic model.

bottom of page